This PHD thesis deals with 3D modelisation in the context of indoor reconstruction of structured environments using LiDAR data. The study aims at automating and improving the pipeline going from the point cloud acquisitions to the 3D reconstruction of building indoors. Indeed, currently, these processes remain mostly manual. The LiDAR data has some specific properties which make the reconstruction challenging (anisotropy, noise, clutters, etc.) and existing methods have a lack of accuracy and their performances depend on the scanned scenes geometry, on the sensor quality and on the acquisition process.

First, the study is oriented towards the point clouds modelisation, one scan at a time. The automatic methods of the state of the art rely on numerous construction hypothesis which yields 3D models relatively far from initial data. The choice has been done to propose a new modelisation method closest to point clouds data, reconstructing only measured areas of each scene and excluding occluded regions. In this objective, we look into the local modelisation process and propose a new normal estimator adapted to structured environments. This tool is integrated to a global modelisation of a scene scanned by a LiDAR device using polygones. This modelisation rely on a joint processing of the range image and the point cloud associated to one scan.

Second, we discuss the registration topic, in order to replace the scans in a global frame. The main objective is to make this process automatic regardless of the scenes geometry, of their initial pose and to obtain good performances for low overlaps. The approaches of the state of the art based on point cloud processing are mostly local and do not seem adapted to structured environments in which local neighborhoods do not carry much information for identification. A new approach adapted to building indoor scenes is proposed to address these issues. Finally, the error resulting from a registration is difficult to measure. Nevertheless, this information is necessary to correct globally a succession of registrations or to fusion the pose information extracted from the registration to other location data provided by external sensors. Some research leads have been explored to estimate the error of a registration and a recent method, based on machine learning, is particularly developped. An adaptation of this method is proposed and evaluated on a synthetic data base. The results emphasize the advantages of the method but also show some critical limitations.